DAY ONE: DPDK book

Since early 1990 network device manufacturer made a lot of innovation in order to increase router speeds. They started from a router node in which everything was computed into the central CPU to reach a situation where the central CPU is less and less used due to a distributed architecture in which lots of action are done in “line cards”.

These progresses have been made thanks to the use of proprietary TCAM (Ternary Content-Addressable Memory) and ASICs (Application-Specific Integrated Circuit) which have been designed to perform table look up and data packets forwarding at high speed.

In early 2000, the Virtualization for x86 computers support has led to lots of innovation into systems domain. Compute virtualization and High-Speed network devices evolution have enabled the Cloud creation.

Later, It appears it was not convenient to manage several isolated network devices each having their own configuration language. Following needs have emerged:

And these desires called for the cloud and SDN technology development.

In Stanford University (US - CA), Clean Slate Research Projects program has been initiated in order to think about how to improve the Internet network architecture. "ETHANE" project was part of this program. Its purpose was to " Design network where connectivity is governed by high-level, global policy". This project is generally known as the first implementation of SDN.

In 2008, a white paper has been proposed by ACM (Association for Computing Machinery) to design a new protocol (OpenFlow) that can program network devices from a network controller.

In 2011, ONF (Open Networking Foundation) has been created to promote SDN Architecture and OpenFlow protocols.

First companies working on SDN have been founded around 2010. (Most of them have now been acquired by main networks or virtualization solution vendors.) In 2007, Martin Casado, who was working on Ethane project has founded Nicira to provide solutions for network virtualization with SDN concept. Nicira has been aquired by vMware in 2012 to develop VMare NSX. In 2016, VMWare also bought PLUMGrid a SDN startup founded in 2013. In 2010, BigSwitch networks has been founded: BigSwitch is proposing a SDN solution. In early 2020, BigSwitch has been acquired by Arista Networks. In 2012, Cisco has created Insieme Networks, a spin-in start-up company working on SDN. In 2013, Cisco take back control on Insieme in order to develop its own SDN solution called ACI (Application Centric Infrastructure). In early 2012, Contrail Systems Inc has been created and aquired at the end of the year by Juniper Networks. In 2013, Alcatel Lucent has created Nuage Networks, a spin-in start-up company working on SDN. Nuage Networks is now an affiliate of Nokia.

The road of SDN development and its history is never straighforward and looks more nuanced than a single storyline might suggest. It’s actually far more complex to be described in a short section. This diagram from [sdn-history] shows developments in programmable networking over the past 20 years, and their chronological relationship to advances in network virtualization.

The concept of SDN, and the term itself, are both very broad and often confusing. There is no real accurate definition of SDN, and vendors usually take it very differently. Initially it was used to in Stanford’s OpenFlow project, and later it has been extended to include a much wider area of technologies. Discussion about each vendor’s exact SDN definition is beyond the scope of this book. But we generally consider that a SDN solution has to provide one to several of following characteristics:

According to [onf-sdn-definition], Software-Defined Networking (SDN) is:

In this diagram, you can see that SDN allows simple high-level policies in the "application layer" to modify the network, because the device level dependency is eliminated to some extent. The network administrator can operate the different vendor-specific devices in the "infrastructure layer" from a single software console - the "control layer". The "controller" in control layer is designed with such a way that it can view the whole network globally. This controller design helps a lot to introduce functionalities or programs as they just needs to talk to the centralized controller, without the need to know all details communicating with each individual device. These details are hidden by the controller from the applications.

Several expectations are behind this new model:

traditionally, A typical network device (e.g. a router) has following planes:

configuration and control plane are located in device’s main processor card, oftenly called "routing engine", or "routing switching engine". The forwarding plane is located in the device’s packet forwarding card, oftenly called "line card".

SDN architecture is built with 3 layers:

In a traditional infrastructure, the route calculation is made on each individual router. each router needs to run one or several routing protocols, through which it exchanges routes with the rest routers in the network, and eventually, based on the route information learned, each router assumes it gains enough knowledge about the network in order to make the forwarding decision. From the network perspective, the control plane is distributed in each individual router, and the end to end routing path is the result of all decisions made by the control plane located on each router.

The control plane on one router may look like this:

In reality, for example, a simplified Juniper MX control plane typical looks like this:

Running a control plane on each router make it very hard to manage, because each individual network device needs to be carefully configured. It requires extensive, vendor-specific experiences and skills to configure the device. The high number of configuration points often make it very challenging to build a robust network. Flexibility is also a recurring hurdle for traditional networks since most routers run proprietary hardware and software.

In contrast, in SDN networking, Control and Configuration functions are gathered into a "SDN controller" which is controlling Network devices. The new architecture intends to provide a completely new way to configure the network. This new Cloud infrastructure brings:

Let’s compare the two architectures:

This SDN infrastructure uses a centralized configuration and control point. route calculation is done centrally in the controller and distributed into each SDN network node. Well the idea looks good and simple, it requires a few foundamental protocols and infrastructures to be implemented before this model can work:

This underlay network infrastructure is playing the same role that the local switch fabric is doing inside a standalone router between the control processor card and lines cards. Based on it, A "overlay" network can be built by the controller, which basically hides underlay network infrastructure details from the applications so they will focus on the high level service implementations. we’ll talk more about "underlay" and "overlay" in the next section.

convenient as it is, this makes the controller the weakest point in the whole model. Think of what will happen if this SDN controller, serving as the "brain", stops working. Everything will be frozen and nothing works as expected, or even worse, some part of the infrastructure continues to run but in an unexpected way, which will very likely trigger bigger issues to other part of the network.

Lots of efforts are done by each SDN solution supplier to solve this weakness. A common and efficient practice is to use clustered architecture to build a highly resilient controller cluster. e.g 3 SDN controllers can load balance and/or backup each other. on failure of one or two, the other one can still make the whole cluster survive, giving the operator longer maintanence windows to fix the problem.

In SDN architecture, each network node is connected to a physical network infrastructure. This physical network which is providing basic connectivity between network nodes is called the "underlay" network infrastructure. sometimes it is also called "fabric", and typically it’s a plane L3 IP network.

very often The underlay needs to separate between different administrative domains (often called "tenants"), switch within the same L2 broadcast domain, route between L2 broadcast domains, provide IP separation via VRFs, and etc. This is implemented in the form of "overlay" networks. The overlay network is a logical network that runs on top of the underlay network. The overlay is formed of tunnels to carry the traffic across the L3 fabric.

Today the industry began to shift in the direction of building L3 data centers and L3 infrastructures, mostly due to the rich features coming from L3 technologies, e.g, ECMP load balancing, flooding control, etc. But the L2 traffic does not disappear and most likely it never will. There are always the desire that a group of network users need to reside in the same L2 network - typically a VLAN. However, In today’s virtualization environment, a user’s VM can be spawned in any compute located anywhere in the L3 cluster. Even if 2 VMs are spawned in the same server, there is often a need to move them around between different servers without changing their networking attributes. These requirements to make a VM always belonging to the "same VLAN" calls for an overlay model over the L3 network. In other words, we need a new mechanism to allow us to tunnel L2 Ethernet domains with different encapsulations over an L3 network.

For example, in SDN node1 we were running VM11 and VM12, they were both serving same sales department and so they were located in same VLAN. because of some administrative requirement, VM12 needs to be moved to another physical SDN node2 which, may be physically located in another rack that is a few router "hops" away. Now we need to ensure not only data packet from VM11 in SDN node1 to be able to reach VM12 in SDN node2, but also they are talking to each other as if they are still in the same VLAN, exactly the same way as before just as if VM12 has never moved. This ability to make the "local" (in same VLAN) traffic to traverse transparently across underlay network infrastructure calls for a packet encapsulation, or "tunneling" mechanism in SDN networks.

Indeed, without such an encapsulation mechanism, traditional segmentation solutions (VLAN, VRF) would have to be provided by the physical infrastructure and implemented up to each SDN node, in order to provide an isolated transportation channel for each customer network connected to the SDN infrastructure.

Encapsulation protocols used in SDN networks have to provide:

Several encapsulation protocols are used into SDN networks:

These encapsulation protocols are providing Overlay connectivity which is required between customers workload connected to the SDN infrastructure. Each SDN node is call a VTEP (Virtual Tunnel End Point) as it is starting and terminating the overlay tunnels.

We’ve seen "openflow" marked as one of the possible interfaces in the "SDN layer" section. Now we’ll introduce the concept of "southbound" and "northbound" interface and other available choices in today’s industry.

The "southbound" interface resides between the controller in "control layer" and network devices in "infrastructure layer". Basically what it does is to provide a means of communication between the 2 layers. Based on the demands and needs, a SDN Controller will dynamically changes the configuration or routing information of network devices. For example, a new VM will advertise a new subnet or host routes when it is spawned in a server, this advertisement will be delivered to SDN controller via a southbound protocol. Accordingly, SDN controller collects all routing updates from the whole SDN cluster through the southbound interfaces, and decides the most current and best route entries, then, it may "reflect" these information to all other network devices or VMs. This ensures all devices having the most uptodate routing information in real time. Among others, examples of the most well-known southbound interfaces in the industry are openflow, OVSDB and XMPP.

OpenFlow is one of the most widely deployed southbound standard from open source community. It first made its appearance in 2008 by Martin Casado at Stanford University. The appearance of OpenFlow was one of the main factors which gave birth to Software Defined Networking.

OpenFlow provides various information for the Controller. It generates the event-based messages in case of port or link changes. The protocol generates a flow based statistic for the forwarding network device and passes it to the controller.

OpenFlow also provides a rich set of protocol specifications for effective communication at the controller and switching element side. Open Flow provides an open source platform for Research Community.

Every physical or virtual OpenFlow-enabled network (data plane) devices in the SDN domain needs to first register with the OpenFlow controller. The registration process is completed via an OpenFlow HELLO packet originating from the OpenFlow device to the SDN controller.

unlike openflow, OVSDB is a southbound API designed to provide additional management or configuration capabilities like networking functions. With OVSDB we can create the virtual switch instances, set the interfaces and connect them to the switches. We can also provide the QoS policy for the interfaces.

The northbound interface provides connectivity between the controller and the network applications running in management plane. As we already discussed that southbound interface has OpenFlow as open source protocol, northbound lacks such type of protocol standards. However with the advancement of technology now we have a wide range of northbound API support like ad-hoc API’s, RESTful APIs etc. The selection of northbound interface usually depends on the programming language used in application development.

with the development of virtualization, SDN technologies and their ecology in recent years, more and more terms and changing of these terms emerge in the networking industry. a lot of confusions have rised, often because of terms are referring different things when they are used in different context. Sometimes the latest term the industry uses is a particular technology such as VNF or a concept such as NFV. Terms rise and fall out of favor as the industry evolves. In recent years the terms such as openstack, NVF/VNF has become the industry’s favorite buzzword. This raises the question - just what is openstack, NVF/VNF and what are the relationships with SDN?

NFV/VNF sounds like new buzzwords, but those technologies have been around for years. according to ETSI:

NFV means "network function virtualization", it stands for an "operation framework for orchestrating and automating VNFs". And VNF means "virtualized network function", such as virtualized routers, firewalls, load balancers, traffic optimizers, IDS or IPS, web application protectors, and so on.

in a nut shell you can think of NVF as a "concept", or "framework" to virtualize certain network functions, while VNF is the implementations of each individual network functions. Among others, firewalls and load balancers are the two most common VNFs in the industry, especially for deployments inside data centers. When you read today’s documents about virtualization technology, you will see the terms in such a pattern like "vXX" (e.g. vSRX, vMX), or "cXX" (e.g. cSRX) very often. that letter v indicates it is a "virtualized" product, while letter c - "containerized" is its container version.

Jointly launched by NASA and Rackspace in 2012, Openstack has rapidly gained popularity in many enterprise data centres. It is one of the most used open source cloud computing platform to support software development and Big Data analytics. OpenStack comprises a set of software modules, e.g, compute, storage & networking modules, which works together to provide an open source choice for building private & public cloud environments. As an IaaS (Infrastructure As A Service) open source implementation ,it provides a wide range of services, from basic service like computing service, storage service, networking service, etc, to advanced services like database, container orchestration and others.

You can think of Openstack as an abstraction layer providing a cloud environment on your promise. with openstack installed in your servers, ,you can spawn a VM, consume and recycle it when you are done, all in seconds. under that abstraction layer, Openstack hides most complexities of automation and orchestration of diverse underlying resources like compute, storage and networking. you could choose Servers, storage, networking devices from your favorite vendors to build the underlying infrastructure, and openstack will "consume" all of them and expose to the user as a pool of common "resources": number of CPUs, RAMs, hard disk spaces, IP addresses, etc. The user does not (need to) care about vendor and brand details.

If we compare openstack with SDN, it’s not hard to see that the two model shares some common features. Both models provide certain level of abstractions, hide the low level hardware details and expose to upper level user applications. the differences are somewhat subtle to describe in just a few words. First off, although there are various distributions from different vendors, they share common core components that is managed by the OpenStack Foundation. SDN is more of a "framework" or an "approach" to manage the network dynamically, which can be implemented with totally different software techniques. Secondly, From the perspective of technical ecological coverage, the ecological aspects of OpenStack are much wider, because networking is just one of its services that is implemented by its Neutron component among it’s other various plugins. SDN, and its ecology, in contrast, mainly focus on the networking. There are also difference in the way that Neutron works comparing with how a typical SDN controller works. OpenStack Neutron focuses on providing network services for virtual machines, containers, physical servers, etc, and provides a unified northbound REST API to users, SDN focuses on configuration and management of forwarding control toward the underlaying network device, it not only provides user-oriented northbound API, but also provides standard southbound API to communicating with various hardware devices.

As we’ve mentioned in previous sections, SDN is a networking scenario which changes the traditional network architecture by bringing all control functionalities to a single location and making centralized decisions. SDN controllers are the brain of SDN architecture, which perform the control decision tasks while routing the packets. Centralized decision capability for routing enhances the network performance. As a result, SDN controller is the core components of any SDN solutions.

While working with SDN architecture, one of the major point of concerns is which controller and solution should be selected for deployment. There are quite a few SDN controller and solutions implementations from various vendors, and every solution has its own pros and cons along with its working domain. In this section we’ll review some of the popular SDN controllers in the market, and the corresponding SDN solutions.

OpenDaylight, aften abbreviated as ODL, is a Java based open source project started from 2013, it was originally led by IBM and Cisco but later hosted under the Linux Foundation. it was the first open source Controller that can support non-OpenFlow southbound protocols, which can make it much easier to be integrated with multiple vendors.

ODL is a modular platform for SDN. It is not a single piece of software. It is a modular platform for integrating multiple plugins and modules under one umbrella There are many plugins and modules built for OpenDaylight. Some are in production, while some are still under development.

Some of the initial SDN controllers had their southbound APIs tightly bound to OpenFlow, But as we can see from the diagram, besides openflow, many other southbound protocols that are available in today’s market are also supported. Examples are NETCONF, OVSDB, SNMP, BGP, etc. Support of these protocols are done in a modular method in the form of different plugins, which are linked dynamically to a central component named "Service Abstraction Layer (SAL)". SAL does translations between the SDN application and the underlaying network equipments. for instance, when it receives a service request from a SDN application, typically via high level API calls (northbound), it understands the API call and translates the request to a language that the underlying network equipments can also understand. That language is one of the southbound protocols.

While this "translation" is transparent to the SDN application, ODL itself needs to know all the details about how to talk to each one of the network devices it supports, their features, capabilities etc. a topology manager module in OLD manages this type of information. What topology manager does is to collect topology related information from various modules and protocols, such as ARP, host tracker, device manager, switch manager, OpenFlow, etc, and based on these info, it visualize the network topology by drawing a diagram dynamically, all the managed devices and how they are connected together will be showed in it.

any topology changes, such as adding new devices, will be updated in the database and reflected immediately in the diagram.

Remember earlier we mentioned that an SDN controller has "global view" of the whole SDN network. In that sense ODL has all necessary visibility and knowledge of the network that can be used to draw the network diagram in realtime.

The Virtualized Cloud Platform (VCP) product from Nuage networks provides a highly scalable policy-based Software-Defined Networking (SDN) platform. It is an enterprise-grade offering that builds on top of the open source Open vSwitch for the data plane along with a feature-rich SDN controller built on open standards.

The Nuage platform uses overlays to provide seamless policy-based networking between Kubernetes Pods and non-Kubernetes environments (VMs and bare metal servers). Nuage’s policy abstraction model is designed with applications in mind and makes it easy to declare fine-grained policies for applications. The platform’s real-time analytics engine enables visibility and security monitoring for Kubernetes applications.

All VCS components can be installed in containers. There are no special hardware requirements.

The Tungsten Fabric (TF), is an open-standard based, proactive overlay SDN solution. It works with existing physical network devices and help address the networking challenges for self-service, automated, and vertically integrated cloud architecture. It also improves scalability through a proactive overlay virtual network technique.

TF controller integrates with most of the popular cloud management systems such as OpenStack, vmware, and Kubernetes. TF’s focus is to provide networking connectivity and functionalities, and enforce user-defined network and security policies to the various of workloads based on different platforms and orchestrators.

Tungsten Fabric’s primary claim to fame is that it is diligently multi-cloud and multi-stack. Today it supports:

TF fits seamlessly into LFN (Linux Foundation Networking) mission to foster open source innovation in the networking space.

The TF system is implemented as a set of nodes running on general-purpose x86 servers. Each node can be implemented as a separate physical server, or VM.

Initially, "Contrail" was a product of a startup company "Contrail system", which was acquired by Juniper Networks in Dec. 2012. It was open sourced in 2013 with a new name "OpenContrail" under the Apache 2.0 license, which means that anyone can use and modify the code of "Opencontrail" system without any obligation to publish or release the modifications. In early 2018, it was rebranded to "Tungsten Fabric" (abbreviated as "TF") as it transitioned into a fully-fledged Linux Foundation project. currently TF is still managed by the Linux Foundation.

Juniper also maintains a commercial version of the Contrail system, and provides commercial support to the payed users. Both The open-source version and commerical version of the Contrail system provide the same full functionalities, features and performances.

TF consists of two main components:

The communication between the controller and vRouters is via XMPP, which is a widely used messaging protocol.

A high level Tungsten Fabric architecture is shown below:

Kernel-based Virtual Machine (KVM) is an open source virtualization technology built into Linux. It provides hardware assist to the virtualization software, using built-in CPU virtualization technology to reduce virtualization overheads (cache, I/O, memory) and improving security.

QEMU is a hosted virtual machine emulator that provides a set of different hardware and device models for the guest machine. For the host, QEMU appears as a regular process scheduled by the standard Linux scheduler, with its own process memory. In the process, QEMU allocates a memory region that the guest sees as physical and executes the virtual machine’s CPU instructions.

With KVM, QEMU can just create a virtual machine with virtual CPUs (vCPUs) that the processor is aware of and runs native-speed instructions. When a special instruction is reached by KVM, like the ones that interacts with the devices or to special memory regions, vCPU pauses and informs QEMU of the cause of pause, allowing hypervisor to react to that event.

LibVirt is an Open Source toolkit to manage virtualization platforms. Libvirt is collection of softwares which allow to manage virtual machines and other virtualization functionality, such as storage and network interface management. LibVirt is proposing to define virtual components in a XML-formatted configurations, that are able to be translated into QEMU command line.

Inter process communication (IPC) is a mechanism which allows processes to communicate with each other and synchronize their actions. The communication between these processes can be considered as a method of cooperation between them.

IPC is used in network virtualization in order to be able to exchange data between different distributed processes of a same application (for example, virtio frontend and backend, contrail vrouter agent and dataplane, etc …​) or between processes of distinct applications (e.g., contrail vrouter and QEMU virtio, virtio and VFIO, and so on)

Two different modes of communication are used for IPC:

Like CPU virtualization, two kinds of methods are used to virtualize network devices:

Software Based Emulation are widely supported but can suffer of poor performance. Hardware assisted Emulation if providing good performance thanks to hardware acceleration, but it requires to use a hardware that supports some specific features.

The following two emulated network devices are provided with QEMU/KVM:

Virtio is an open specification for virtual machines' data I/O communication, offering a straightforward, efficient, standard and extensible mechanism for virtual devices, rather than boutique per-environment or per-OS mechanisms. It uses the fact that the guest can share memory with the host for I/O to implement that.

Virtio was developed as a standardized open interface for virtual machines (VMs) to access simplified devices such as block devices and network adaptors.

KVM guests usually have access to software based emulated NIC device (either para-virtualized devices with virtio or traditional emulated devices). On host machines which have Intel VT-d or AMD IOMMU hardware support, another option is possible. PCI devices may be assigned directly to the guest, allowing the device to be used with minimal performance overhead.

Assigned devices are physical devices that are exposed to the virtual machine. This method is also known as passthrough.

The VT-d or AMD IOMMU extensions must be enabled in BIOS in order to be able to perform for device Direct Assignment:

Two methods are supported:

With VFIO the Physical device is exposed to the host user space memory and is made visible from the guest VM it has been assigned.

Single Root I/O Virtualization (SR-IOV) specification is defined by the PCI-SIG (PCI Special Interest Group). This is a PCI Express (PCI-e) that extends a single physical PCI function to share its PCI resources as separate virtual functions (VFs).

The physical function contains the SR-IOV capability structure and manages the SR-IOV functionality (it can be used to configure and control a PCIe device).

A single physical port (root port) presents multiple, separate virtual devices as unique PCI device functions (up to 256 virtual functions – depends on device capabilities).

Each virtual device may have its own unique PCI configuration space, memory-mapped registers, and individual MSI-based interrupts. Unlike a physical function, a virtual function can only configure its own behavior. Each virtual function can be directly connected to a virtual machine via PCI device assignment (passthrough mode).

SR-IOV improves network device performance for each virtual machine as it can share a single physical device between several virtual machines using device direct I/O assignment method.

With SR-IOV, each VM has a direct access to the physical network using the assigned virtual function interface allocated to each. They can communicate altogether using the Virtual Ethernet Bridge provided by the NIC card. A virtual switch can also use SRIOV to get access to the physical network. VM using SRIOV assigned virtual function device has a direct access to the physical network and are not connected to any intermediate virtual network switch or router.

Following command can be used to check whether SR-IOV is supported or not on a physical NIC card:

$ lspci -s <NIC_BDF> -vvv | grep -i "Single Root I/O Virtualization"

VirtIO is bringing lots of flexibility. VirtIO is offering a standardized driver which is fully independent of the hardware used on the physical platform hosting VM instances.

When virtio connectivity is used VM can be easily migrated from one host to another using "live migration" feature. When SRIOV is use, this live migration is not an easy task and is not really possible to achieve.

Indeed, network driver used by VM depends on used hardware on the bare metal node which are hosting them. In order to make VM migration from one bare metal node to another, both nodes must at least to use same hardware NIC model. But when SRIOV is used VM connectivity is having barely the same performance has a real physical NIC, whereas with VirtIO, performance could be poor.

Also, SRIOV, providing a direct access to the physical NIC is making host virtual network nodes (virtual router/switch) used by SDN solution totally blind about VM using such connectivity. Local traffic switching between VM connected on a same SRIOV physical card is achieve by the Virtual Ethernet bridge proposed by SRIOV. Communication between VM connected onto distinct SRIOV physical ports must rely on physical network.

SDN vswitch/vrouter usage is very limited when SRIOV is used. Indeed, packet switching between VMs which are using VFs of a same SR-IOV physical port are using the physical Virtual Ethernet Bridge hosted in the physical NIC.

Only some few use cases are relevant, which are:

With SRIOV we are getting high performance but with poor flexibility and no network virtualization features. With VirtIO we are getting a high level of network virtualization suitable for SDN, which is very flexible with poor performances.

For SDN use cases, we need network virtualization features and performance. DPDK will bring both.

DPDK is dedicating one (or more) CPU to one (or more) thread that are continuously polling a one (or more) DPDK NIC RX queue. CPU on which a DPDK polling thread is started will be loaded at 100% whatever there some packets to process or not, as no interrupt mechanism is used in DPDK to warn the DPDK application that a packet has been received.

Using DPDK library API, physical NIC packets will be made available into user space memory in which the DPDK application is running. So, when DPDK is used there is no user space to kernel space context switching and it saves lots of CPU cycles. Also, the host memory is using large continuous memory area, the huge pages, which allow large data transfers and avoid high data fragmentation in memory which would require a higher memory management effort at the application level. Such a fragmentation would also cost some precious CPU cycles.

Hence, most of the CPU cycles of DPDK pinned CPU are used for polling and processing packets delivered by the physical NIC in DPDK queues. As a result, the packet forwarding task can be processed at a very high speed. If one CPU is not powerful enough to manage incoming packets that are hitting the physical NIC at a very high rate; we can allocate an additional one to the DPDK application in order to increase its packet processing capacity.

A DPDK application is a multi-thread program that is using DPDK library to process network data. In order to scale, we can start several packet polling and processing threads (each one pinned on a dedicated CPU) that are running in parallel.

3 main components are involved into a DPDK application:

Data Plane Development Kit (DPDK) is a set of data plane libraries and network interface controller drivers for fast packet processing, currently managed as an open-source project under the Linux Foundation.

The main goal of the DPDK is to provide a simple, complete framework for fast packet processing in data plane applications.

The framework creates a set of libraries for specific environments through the creation of an Environment Abstraction Layer (EAL), which may be specific to a mode of the Intel® architecture (32-bit or 64-bit), Linux* user space compilers or a specific platform.

These environments are created through the use of make files and configuration files. Once the EAL library is created, the user may link with the library to create their own applications.

The DPDK implements a "run to completion model" for packet processing, where all resources must be allocated prior to calling Data Plane applications, running as execution units on logical processing cores.

The model does not support a scheduler and all devices are accessed by polling. The primary reason for not using interrupts is the performance overhead imposed by interrupt processing.

For more information please refer to dpdk.org documents http://dpdk.org/doc/guides/prog_guide/index.html

DPDK is a set of programing libraries that can be used to create an application that needs to process network packets at a high speed. DPDK is proposing following functions:

In order to reduce Linux user to kernel space context switching all these functions are made available by DPDK into the user space where applications are running. User applications using DPDK libraries have a direct access to the NIC cards, without passing through a NIC Kernel driver as it is required when DPDK is not used.

DPDK is allowing to build user-space multi-thread network application using the POSIX thread (pthread) library.

DPDK is a framework which is made of several libraries:

Ethdev library exposes APIs to use the networking functions of DPDK NIC devices. The bottom half part of ethdev is implemented by NIC PMD drivers. Thus some features may not be implemented.

Poll Mode ethernet Drivers (PMDs) are a key component for DPDK. These PMDs by-pass the kernel and are providing a direct access to the Network Interface Cards (NIC) used with DPDK.

Linux user space device enablers (UIO or VFIO) are provided by Linux Kernel and are required to run DPDK. They are allowing to discover and expose PCI devices information and address space through the /sys directory tree.

DPDK libraries are allowing kernel-bypass application development:

The diagram below is providing an overview of DPDK libraries.

Only few libraries have been described in this diagram: Set of libraries is enriched at each new DPDK release (cf: https://www.dpdk.org/).

The Environment Abstraction Layer (EAL) is responsible to provide access to low-level resources such as hardware and memory space. It provides a generic interface that hides the environment specifics from the applications and libraries. The EAL performs physical memory allocation using mmap() in hugetlbfs (using huge page sizes to increase performance).

Provided services by EAL are:

DPDK has a highly optimized memory manager. DPDK works on a group of fixed size objects called a mempool. Every one of them are pre-allocated. DPDK does not encourage dynamic allocations because it consumes a lot of CPU cycles and it is a speed killer.

DPDK stores incoming packets into mbufs (memory buffers). DPDK pre-allocates a set of mbufs and keeps it in a pool called mempool.

DPDK makes use of mempools each time it needs to allocate a mbuf where packets are stored. Instead of allocating a single mbuf, DPDK do a bulk allocation, or bulk free once packets are consumed. By doing this, packets to be processed (mbufs) are already in cache memory. Therefore, DPDK is very cache friendly.

Mempool has further optimizations. It is very cache friendly. Everything is aligned to the cache and has a some mbufs allocated for each DPDK thread or lcore. Each mempool are also bound with rings which are referencing mbufs containing packets stored into mempool.

Each ring is a highly optimized lockless ring. It can be used by several lcores in a multi-producer/multi-consumer kind of scenario without locks. By avoiding locks, DPDK gets large performance gains, as data structures locking is also a speed killer.

Network Data are stored in compute central memory (in huge page area).

DPDK uses message buffers known as mbufs to store packet data into the host memory. These mbufs are stored in memory pools known as mempools.

mbufs are storing DPDK NIC incoming and outgoing packets which have to be processed by the DPDK application.

DPDK queues are not storing the packets but a pointer onto the real packet. It avoids performing a data transfer that would be needed when packets have to be forward from a DPDK NIC to another.

Packets are not moved from one queue to another, but these are descriptors (pointers) that are moving from one queue to another.

Descriptors are set up as a ring. A ring is a circular array of descriptors. Each ring describes a single direction DPDK NIC queue. Each DPDK NIC queue is made up of 2 rings (1 per direction: 1 RX ring, 1 TX ring).

Each descriptor points onto a packet that has been received (RX ring) or that is going to be transmitted (TX ring).

The more descriptors RX/TX rings are containing, the more memory size will be required in each mempool (number of mbufs) to store data.

DPDK application is only processing packets that are exposed in user space host OS memory.
DPDK rings are an abstraction of the real NIC queues: DPDK is using DMA to keep synchronized at anytime between the NIC hardware queues and its DPDK representation in the host memory.

Multithreading is the ability of a CPU (single core in a multi-core processor architecture) to provide multiple threads of execution concurrent. In a multithreaded application, the threads share some CPU resources memory:

A single Linux process can contain multiple threads, all of which are executing the same program. These threads share the same global memory (data and heap segments), but each thread has its own stack (local variables).

Linux pThreads (POSIX threads) is a C library which contains a set functions that are allowing to manage threads into an application. DPDK is using Linux pThreads library.

DPDK is using threads that are designed as "lcore”. A “lcore" refers to an EAL thread, which is really a Linux pthread, which is running onto a single processor execution unit.

Lcores are not sharing CPU units. Nevertheless, if the host processor supports hyperthreading, a core may include several lcores or threads.

lcores are used to run DPDK application packet processing threads. Several packet processing models are proposed by DPDK. The simplest one is the Run-To-Completion model.

Run-to-Completion, is using a single thread (lcore) for end to end packet processing (packet polling, processing and forwarding).

A complex application is typically split across multiple cores, with cores communicating through Software queues.

Packet Framework facilitates the creation of pipelines. Each pipeling thread is assigned to a CPU and is using software queues like output or/and input ports.

For instance, Contrail DPDK vRouter is using such a model for GRE encapsulated packet processing.

It is possible to create Control Threads. Those threads can be used for management/infrastructure tasks and are used internally by DPDK for multi process support and interrupt handling.

DPDK service cores enables a dynamic way of performing work on DPDK lcores. Service core support is built into the EAL, and an API is provided to optionally allow applications to control how the service cores are used at runtime.

PCI devices have a set of registers referred to as configuration space for devices. These configuration space registers are mapped to host memory locations.

When a PCI device is enabled, the system’s device drivers (by writing configuration commands to the PCI controller) programs the Base Address Registers (BAR) to inform the PCI device of its address mapping. Next, the host operating system is able to address this PCI device.

With usual Linux NIC Kernel, both NIC configuration and Packet processing is done in Kernel Space. User applications which have to establish a TCP connection or send a UDP packet is using the sockets API, exposed by libc library.

Linux Packet Processing with sockets API is requiring following operations which are costly:

With usual Linux Drivers most of operations are occurring in Kernel modes and are requiring lots of user space to kernel space context switching and interruption mechanisms. The heavy context switching usage is costing lots of CPU cycles and is a limiting the numbers of packets that a CPU is able to process. Such drivers are not able to perform packet processing at expected high speed, especially when 10/40/100G Ethernet generation cards are used on a Linux System.

A Poll Mode Driver consists of APIs, running in user space, to configure the devices and their respective queues. In addition, a PMD accesses the RX and TX descriptors directly without any interrupts (with the exception of Link Status Change interrupts) to quickly receive, process and deliver packets in the user’s application.

Poll Mode drivers are involved in NIC configuration. They are exposing NIC configuration registers into host memory area which is directly reachable from user space.

In short, Poll Mode Drivers are user space pthreads which:

Hence user applications can configure directly the NIC cards they are using from Linux user space where they are running.

A first configuration phase is using Poll Mode Drivers and DPDK library to configure DPDK rings buffers into Linux user space. Next, incoming packets will be automatically transferred with DMA (Direct Memory Access) mechanism from NIC physical RX queues in NIC memory to DPDK RX rings buffer in host memory. DMA (Direct Memory Access) is also used to transfer outgoing packets from DPDK TX rings buffer in host memory to NIC physical TX queues in NIC memory. DMA offloads expensive memory operations, such as large copies or scatter-gather operations, from the CPU.

Direct Memory Access (DMA) allows PCI devices to read (write) data from (to) memory without CPU intervention. This is a fundamental requirement for high performance devices.

DMA is a mechanism that is using a specific hardware controller to manage read and write operations into the main system memory (RAM: Random Access Memory). This mechanism is totally independent of the central processing unit (CPU) and does not consume any CPU resource. A DMA transfer is used to manage data transfer. DMA transfer is triggered by the CPU and is working in background using the specific hardware resource (DMA controller).

DPDK rings and NIC buffers are synchronized with DMA. Thanks to this synchronization mechanism, DPDK application can access transparently to NIC packets in user space reading or writing data in DPDK rings.

Input–Output Memory Management Unit (IOMMU) is a memory management unit (MMU) that connects a Direct Memory Access (DMA) capable I/O bus to the main memory.

In Virtualization, an IOMMU is re-mapping the addresses accessed by the hardware into a similar translation table that is used to map guest virtual machine address memory to host-physical addresses memory.

IOMMU provides a short path for device to get access only to a well scoped physical device memory area which corresponds to a given guest virtual machine memory. IOMMU helps to prevent DMA attacks that could be originated by malicious devices. IOMMU provides DMA and interrupt remapping facilities to ensure I/O devices behave within the boundaries they’ve been allotted.

Intel has published a specification for IOMMU technology as Virtualization Technology for Directed I/O, abbreviated as VT-d.

In order to get IOMMU enabled:

GRUB configuration example with IOMMU Passthrough enabled:

DPDK Library includes Poll Mode Drivers (PMDs) for physical and emulated Ethernet controllers which are designed to work without asynchronous, interrupt-based signaling mechanisms.

Lots of other NIC are supported by DPDK (cf http://doc.dpdk.org/guides/nics/overview.html).

Different PMDs may require different kernel drivers in order to work properly (cf Linux User space device enablers). Depending on the PMD being used, a corresponding kernel driver should be loaded and bound to the network ports.

This is also preferable that each NIC has been flashed with the latest version of NVM/firmware.

Linux kernel version 2.6 introduced the User Space IO (UIO) loadable module. UIO is a kernel-bypass mechanism which provides an API that enables user space handling of legacy interrupts (INTx).

UIO has some limitations:

UIO only supports legacy interrupts so it is not usable with SR-IOV and virtual hosts which require MSI/MSI-X interrupts.

Despite these limitations, UIO is well suited for use in virtual machines, where direct IOMMU access is not available. In such a situation, a guest instance user space process is not isolated from other processes in the same instance. But the hypervisor can isolate any guest instance from others or hypervisor host processes using IOMMU.

Currently, two UIO modules are supported by DPDK:

DPDK specific UIO Kernel module is loaded with insmod command after UIO module has been loaded:

While a single command is needed to load Linux Generic UIO module:

DPDK specific UIO module could be preferred in some situation to Linux Generic UIO module (cf: https://doc.dpdk.org/guides/linux_gsg/linux_drivers.html)

Virtual Function I/O (VFIO) kernel infrastructure was introduced in Linux version 3.6.

VFIO provides a user space driver development framework allowing user space applications to interact directly with hardware devices by mapping the I/O space directly to the application’s memory.

VFIO is a framework for building user space drivers that provides:

VFIO exposes APIs which allow to:

VFIO driver is an IOMMU/device agnostic framework for exposing direct device access to user space, in a secure, IOMMU protected environment. For bare-metal environments, VFIO is the preferred framework for Linux kernel-bypass. It operates with the Linux kernel’s IO.

MMU subsystem is used to place devices into IOMMU groups. User space processes can open these IOMMU groups and register memory with the IOMMU for DMA access using VFIO ioctl calls. VFIO also provides the ability to allocate and manage message-signaled interrupt vectors.

A single command is needed to load VFIO module:

Despite VFIO has been created to work with IOMMU, VFIO can be also be used without (this is just as unsafe as using UIO).

VFIO is generally the preferred Linux user space device enabler to be used because it supports IOMMU to protect host memory. When a real hardware PCI device is attached to host system and IOMMU is used with VFIO, all the reads/writes of that device done in user space by the DPDK application will be protected by the host IOMMU.

But there some is few exceptions. Below is Intel recommendation for the choice of the Kernel driver to be used with DPDK:

https://software.intel.com/content/www/us/en/develop/articles/memory-in-dpdk-part-2-deep-dive-into-iova.html

NUMA means Non-Uniform Memory Access systems

A traditional server has a single CPU, a single RAM and a single RAM controller.

A RAM can be made of several DIMM banks in several sockets, all being associated to the CPU. When the CPU needs access to data in RAM, it requests it to its RAM controller.

Recent servers can have multiple CPUs, each one having its own RAM and its own RAM controller. Such systems are called NUMA systems, or Non-Uniform Memory Access. For example, in a server with 2 CPUs, each one can be a separate NUMA: NUMA0 and NUMA1.

NUMA nodes architecture.

When CPU0 needs to access data located in RAM0, it will go through its local RAM controller 0. Same thing happens for CPU1.

When CPU0 needs to access data located in the other RAM1, the first (local) controller 0 has to go through the second (or remote) RAM controller 1 which will access the (remote) data in RAM 1. Data will use an internal connection between the 2 CPUs called QPI, or Quick Path Interconnect, which is typically of a high enough capacity to avoid being a bottleneck, typically 1 or 2 times 25GBps (400 Gbps). For example, the Intel Xeon E5 has 2 CPUs with 2 QPI links between them; Intel Xeon E7 has 4 CPUs, with a single QPI between pairs of CPUs.

The fastest RAM that the CPU has access to is the register, which is inside the CPU and reserved to it.

Beyond the register, the CPU has access to cached memory, which is a special memory based on higher performance hardware.

Cached memories are shared between the cores of a single CPU. Typical characteristics of memory cache are:

If the CPU needs to access data that is in the main RAM, it has to use its RAM controller.

Access to RAM takes typically 170 CPU cycles (the green line in the diagram). Access to the remote RAM through the remote RAM controller typically adds 200 cycles (the red line in the diagram), meaning RAM latency is roughly doubled.

When data needed by the CPU is located both in the local and in the remote RAM with no particular structure, latency to access data can be unpredictable and unstable.

A single physical CPU core with hyper-threading appears as two logical CPUs to an operating system.

While the operating system sees two CPUs for each core, the actual CPU hardware only has a single set of execution resources for each core.

Hyper-threading allows the two logical CPU cores to share physical execution resources.

The sharing of resources allows two logical processors to work with each other more efficiently and allows a logical processor to borrow resources from a stalled logical core (assuming both logical cores are associated with the same physical core). Hyper-threading can help speed processing up, but it’s nowhere near as good as having actual additional cores.

Memory is managed in blocks known as pages. On most systems, a page is 4KB. 1MB of memory is equal to 256 pages; 1GB of memory is 256,000 pages, etc. CPUs have a built-in memory management unit that manages a list of these pages in hardware.

The Translation Lookaside Buffer (TLB) is a small hardware cache of virtual-to-physical page mappings. If the virtual address passed in a hardware instruction can be found in the TLB, the mapping can be determined quickly. If not, a TLB miss occurs, and the system falls back to slower, software-based address translation. This results in performance issues. Since the size of the TLB is fixed, the only way to reduce the chance of a TLB miss is to increase the page size.

Virtual memory address lookup slows down when the number of entries increases.

A huge page is a memory page that is larger than 4Ki. In x86_64 architecture, in addition to standard 4KB memory page size, two larger page sizes are available: 2MB and 1GB.

Contrail DPDK vrouter can use both or only one huge page size.

An Operating System is using a scheduler to place each single process and/or threads it has to run onto one CPUs offered by a host.

There are two kinds of scheduling, cooperative and preemptive. By default, Linux scheduler is using a cooperative mode.

In order to get a CPU booked for a subset of tasks, we have to inform the Operating System scheduler not to use these CPUs for all the tasks it has to run.

These CPUs are told: "isolated" because they are no more used by the OS to process all tasks. In order to get a CPU isolated several mechanisms can be used:

Isolation and pinning are two complementary mechanisms that are proposed by Linux OS:

Tasks to be run by an operating system must be spread across available CPUs. These tasks in a multi-threading environment are often made of several processes which are also made of several threads.

A thread’s CPU affinity mask determines the set of CPUs on which it is eligible to run.

Linux systemd is a software suite that provides an array of system components for Linux operating systems. Its primary component is an init system used to bootstrap user space and manage user processes.

CPUAffinity parameter restricts all processes spawned by systemd to the list of cores defined by the affinity mask.

taskset is used to set or retrieve the CPU affinity of a running process given its PID or to launch a new COMMAND with a given CPU affinity.

We can retrieve the CPU affinity of an existing task:

Or set it:

Virtual IOMMU (vIOMMU) is allowing to emulate IOMMU for guest VMs.

vIOMMU has the following characteristics:

In order to get a virtual device working with a virtual IOMMU we have to:

Following mechanisms can be used to create vIOMMU memory mappings:

The integration between the virtual IOMMU and any user space network application like DPDK is usually done through the VFIO driver. This driver will perform device isolation and automatically add the memory (IOVA -to GPA) mappings to the virtual IOMMU.

The use of hugepages memory in DPDK contributes to optimize TLB lookups, since a fewer number of memory pages can cover the same amount of memory. Consequently, the number of Device TLB synchronization messages drop dramatically. Hence, the performance penalty TLB lookups is lowered.

Cf: https://www.redhat.com/en/blog/journey-vhost-users-realm

https://wiki.qemu.org/Features/VT-d

Virtio-pmd driver, is a DPDK driver, built on the Poll Mode Driver abstraction, that implements the virtio protocol.

Vhost user protocol moves the virtio ring from kernel all the way to userspace. The ring is shared between the guest and DPDK application. QEMU sets up this ring as a control plane using unix sockets.

If the both the host server guest virtual machine are DPDK there are no VMExits in the host for guest packets processing. Guest virtual machine uses virtio-net PMD driver and performs packets polling. So. There is nothing running in kernel here, so there are no system calls. Since both system calls and VM Exits are avoided, the performance boosts significantly. It will be an order higher.

When a DPDK application is running into a guest Virtual Machine, a mechanism has to be used to expose one of the host physical NIC to this guest in order it gets access to the physical network.

IOMMU protects host memory against malicious or bug writes which can corrupt host memory at any time. But, when a physical device is assigned to a guest virtual machine without vIOMMU usage, the guest memory address space is totally exposed to the hardware PCI device.

A PCI device can be assigned to a guest in order to be used by a guest DPDK application. By leveraging VFIO driver in the host kernel we provide a direct access to an assigned physical NIC from this guest protected with IOMMU.

Next, by leveraging VFIO driver in the guest kernel we provide a direct access to the assigned physical from this guest user space. vIOMMU is providing a secure mechanism to manage DMA transfer between an assigned physical hardware and hosted guest virtual instance memory area.

This use case is almost the same as PCI passthrough. VFIO and IOMMU are used to expose a SRIOV virtual function directly to a guest VM.

An additional Physical function driver which is vendor specific is used to manage the virtual function creation on the physical NIC. This driver is used by a Virtual Machine Manager (like libvirt) to create the virtual function before the virtual instance is spawned.

Physical incoming packets are directly copied in guest memory without involving the host server. SR-IOV only allow to share a physical NIC between several guests but does not change the packet processing path provided by PCI passthrough.

With virtio full hardware offloading, both the virtio data plane and virtio control plane are offloaded to the NIC hardware. The physical NIC must support:

Hence once the guest memory is mapped with the NIC using virtio physical device passthrough, the guest communicates directly with the NIC via PCI without involving any specific drivers in the host kernel.

Guest VM packet processing is directly performed in NIC hardware like but presented to the guest instance like a regular virtio emulated interface. Guest VM does not make any difference between a virtio emulated interface and an assigned physical virtio NIC, as they are exposed with the same virtio driver frontend in the guest.

virtio device passthrough

Virtio device passthrough can be implemented onto a NIC which is supporting or not SR-IOV.

Like other physical device assignment technics presented in this book, VFIO and IOMMU are used to present the physical device NIC into the guest VM user space.

Hence, such a virtio physical NIC can be used by a DPDK application running into a virtual instance. But, like other virtio device passthrough has also the same limitations for SDN. As the host operating system is totally by passed by this mechanism, we cannot interconnect instances using such NIC interface with a SDN virtual router or switch.

The main advantage of Virtio device passthrough is the flexibility it provides for a virtual instance to use transparently either a real physical interface or an emulated one. It offers an Open public specification, which provide device fully independent of any specific vendor.

Virtio full HW offloading, can support live migration thanks to virtio, which is not possible to achieve without any specific implementation with SR-IOV.

But in order to be able to support such a feature, latest virtio specifications (1.1 version) must be implemented onto both QEMU and the NIC hardware used on the cloud infrastructure.

Like full hardware offloading, virtual Data Dath Acceleration (vDPA) aims to:

vDPA is presenting a generic control plane through a software piece which provides an abstraction layer on top of physical NIC.

Like Virtio full hardware offloading, vDPA build a direct data path between the gest network interface and the physical NIC, using the virtio ring layout. But for the control path a generic vDPA driver (mediation driver) is used to translate the vendor NIC driver/control-plane to the VirtIO control plane, in order to allow each NIC vendor to keep using its own driver.

It allows NIC vendors to support virtio ring layout at smaller effort keeping wire speed performance on the data plane.

virtio datapath acceleration

vDPA is requiring a vendor specific "mediation device driver" to be loaded in the host operating system.

A NIC card generation commonly named "smart NIC" are highly customizable thanks to the last evolution provided by some new capabilities (FPGA, P4).

It makes possible to envisage SDN vSwitch/vRouter dataplane function to be moved into the NIC card keeping only the controle plane function into the host operating system.

For Contrail solution, this is made by offloading several Contrail vRouter tables including:

It allows to accelerate lookups and forwarding actions that are directly performed into the NIC.

SDN packet processing is fully done into the NIC card, no more host CPU processing is involved in packet processing.

Two implementations are proposed by Metronome:

SRIOV + SmartNIC:

vDPA + Smart NIC:

Berkeley Packet Filter (BPF) was designed for capturing and filtering network packets that matched specific rules. In last years extended BPF (eBPF) has been designed to take advantage of new hardware (64 bits usage for intance). An eBPF program is "attached" to a designated code path in the kernel.

eXpress Data Path (XDP), uses eBPF to achieve high-performance packet processing by running eBPF programs at the lowest level of the network stack, immediately after a packet is received. XDP.

XDP support is made available in the Linux Kernel since version 4.8, while eBPF is supported in the Linux Kernel since version 3.18.

XDP requires:

XDP packet processor performs:

BFP programs:

BFP program main actions are :

XDP is also able to offload an eBPF program to a NIC card which supports it, reducing the CPU load.

XDP and eBPF does not require:

PS: eBPF rules are also supported in DPDK application.

https://www.redhat.com/en/blog/using-express-data-path-xdp-red-hat-enterprise-linux-8